prefrontalos

prefrontal-systems/prefrontalos

3.1

If you are the rightful owner of prefrontalos and would like to certify it and/or have it hosted online, please leave a comment on the right or send an email to dayong@mcphub.com.

PrefrontalOS is an MCP server designed to help AI assistants recognize and break out of failure patterns by detecting loops, injecting directives, and learning what works.

PrefrontalOS

Anastrophe (ἀναστροφή) = conduct, behavior in ancient Greek Following the mnemex naming convention: memory + codex → conduct + codex

An MCP server designed to help AI assistants recognize and break out of failure patterns by:

  1. Detecting loops - monitoring tool calls and command patterns
  2. Injecting directives - providing timely reminders from CLAUDE.md/AI_CREDO.md
  3. Learning what works - tracking effectiveness via mnemex integration

The Problem

AI assistants (like me) can get stuck in trial-and-error loops:

  • Making similar attempts with slight variations
  • Guessing instead of using verification tools
  • Ignoring documented best practices
  • Not recognizing repeated failures

Real example (Oct 23, 2025): Spent 5+ commits manually editing Python file formatting instead of running black file.py.

The Solution

A three-layer system:

┌──────────────────────────────────────────────┐
│ 1. Pattern Detection                         │
│    - Tool call history                       │
│    - Similar commands with variations        │
│    - Failed attempt counter                  │
└──────────────────────────────────────────────┘
              ↓
┌──────────────────────────────────────────────┐
│ 2. Directive Injection                       │
│    - Pull from ~/.claude/CLAUDE.md           │
│    - Context-aware reminders                 │
│    - Lightweight alerts (not blocking)       │
└──────────────────────────────────────────────┘
              ↓
┌──────────────────────────────────────────────┐
│ 3. Learning via Mnemex                       │
│    - Write findings to mnemex graph          │
│    - Track what reminders work               │
│    - Persist across Claude sessions          │
└──────────────────────────────────────────────┘

Architecture Goals

Persistence Across Sessions

  • Claude doesn't persist, but the MCP server and mnemex do
  • System learns which patterns/reminders are effective over time
  • Each new Claude session benefits from past learnings

Incremental Gains

  • Not trying for perfection - "pobody's nerfect"
  • 30% success rate is a win - better than 0%
  • Progress over perfection - catch some loops, improve over time

Effectiveness Tracking

Pattern: Manual formatting edits (3+ times)
Reminder: "Use `black file.py` instead of guessing"
Effectiveness: 4/5 (worked 80% of the time)
Context: Python repos with Black configured

Documented Failure Patterns

See for detailed patterns from real debugging sessions.

Case Studies

  • - Oct 23, 2025

Installation

# Development installation
cd behavior-mcp
uv pip install -e ".[dev]"

# Run the server
prefrontalos

# Or using Python module
python -m prefrontalos.server

Development

Status: Planning/Design Phase Package name: prefrontalos Entry point: prefrontalos command

See for detailed implementation plan.

Quick start:

# Run tests
uv run pytest

# Format code
uv run black src/ tests/

# Type check
uv run mypy src/

Contributing

This is a collaborative design exercise. All insights, patterns, and effectiveness data welcome.

Core Philosophy: Any step that helps AI get better is a win. We're building a safety net that catches some failures, then improving it over time.